a statistical model of singapore housing prices · rates, housing policies, etc.) • noise...
TRANSCRIPT
Singapore Actuarial Society Health and Retirement Conference 2018
A Statistical Model of Singapore Housing Prices
Authors:
Minhao Leong FSA, FRM, CAIA
Leonardi Tjokro FSA, CERA
Background of Speakers
Minhao Leong
Professional Profile:• Actuarial professional with experience across insurance
and credit risk analytics• Professional involvement with the predictive analytics
functions of the SOA and the IoAA
Education:• MS Analytics Candidate, Georgia Tech• BCom Finance (Hons), USYD• BCom Actuarial Studies, UNSW
Interests:• Statistical modelling projects in FMCG, sports
predictions and real estate finance
Leonardi Tjokro
Professional Profile:• Pricing actuary with strategy experience across
distribution channels and agency compensation• Currently a senior strategy manager at Manulife
Indonesia overseeing distribution projects
Education:• FSA, SOA• CERA, SOA• BCom Actuarial Studies, UNSW
Interests:• Investments and digital analytics
Overview of today’s presentation
Singapore housing prices
Macro level housing model
Overview of global residential markets
Framework for modeling housing prices• Current and future stock of residential properties• Trend - Demographic variables (population growth, real
income, migration, etc.)• Cyclical – Economic variables (GDP position, interest
rates, housing policies, etc.)• Noise – Speculative buying and currency driven demand
Use cases:• Understanding the current state of the Singapore
housing market• Evaluating the implications of government policies• Forward looking projection of broad housing market
direction
Transactional level housing model
Factors affecting transactional housing prices:• Geographical variables (ie postal district, distances to
key landmarks)• Individual property characteristic variables (ie floor,
age)• Macroeconomic variables (ie interest rates, GDP per
capita)• Demographic variables (ie population growth, size of
middle class)
Pricing approaches:• Traditional – Linear models for risk based pricing• Progressive – Penalised linear models, decision trees• Radical – Statistical learning approaches
Use cases:• Providing a theoretical ‘arbitrage price’ for evaluating
fair real estate valuations• Understanding the key drivers of individual real estate
prices
Singapore Actuarial Society Health and Retirement Conference 2018
Section 1: Modelling house prices – A macroeconomic perspective
Comparing different housing markets across the globe
• Firm monetary and fiscal policies
• Household income: High
• Population growth: Above average
• GDP growth: Stable/Accelerating
• Loose monetary and fiscal policies after GFC
• Household income: Average
• Population growth: Below average
• GDP growth: Decelerating
• Tight monetary and fiscal policies after GFC
• Household income: Average
• Population growth: Average
• GDP growth: Stable
• Tight monetary and fiscal policies after GFC
• Household income: Below average
• Population growth: Mixed
• GDP growth: Decelerating
Boom-Busters
(Eurozone)
Stabilizers
(US, UK)
High-Risers
(Australia, NZ, Canada)
Deflators
(Germany, Switzerland,
Japan)
Differences in housing markets due to supply, demographic and economic factors
Source: Savills World Research; World Residential Markets
An introduction to Singapore’s housing markets
Tracing the evolution of Singapore’s housing market “From Third World to First”
1997: Asian Financial Crisis
1996: Implementation of anti-speculation measures
2007-2008: Global Financial Crisis
Providing a framework for modeling housing prices
Factors Considered Used Supporting explanation
1) Supply variables
Current Supply Determines the number of units potentially available at the present day. A statistically significant driver of price.
Future Supply Future supply weakly drives current housing prices.
2) Demographic variables
Population trend Size of working population has a statistically significant impact on housing prices, more specifically age group 20-39 and 60 over.
Net migration Another important driver of demand for residential properties.
Income trend at each decile Tracks the housing affordability index of the population.
3) Economic variables
Interest Rate Sets the mortgage payments, which directly impact the propensity of taking a housing loan.
GDP position Measures consumer confidence in purchasing a house and banks confidence in extending loans. An indicator of the state of the economy.
Cooling measures (Indicators)
Captures their immediate impact on housing prices.
STI Index position Similar to GDP position.
4) Speculation variables
Purchase of 2nd homes Proxy for domestic speculative buying
Currency pairs Proxy for foreign speculative buying
What drives Singapore’s housing prices?
Making our case for a focus on demand-related variables (demographic, economic and speculation variables)
Decomposing housing prices: Focus on demand-related variables
Demographic factors drive the long term trend; economic cycles and speculative behaviors cause the temporary fluctuations
Decomposing housing prices: Demographic variables
Understanding the demographic drivers of Singapore housing prices
40%
45%
50%
55%
60%
65%
70%
-
1
2
3
4
5
6
7
1975 1980 1985 1990 1995 2000 2005 2010 2015
Mill
ion
s
Population trend
Working (age 20-64) Not working Working %
-15%
0%
15%
0
2
4
6
8
10
12
14
2000 2005 2010 2015
Tho
usa
nd
s
Income trend (monthly income)
10th decile 5th decile
2nd decile 10th decile (real %)
5th decile (real %) 2nd decile (real %)
Housing activities move in the same direction with that of the overall economy, but tend to have a larger volatility
Decomposing housing prices: Economic variables
-1
-0.5
0
0.5
1
1.5
2
1980 1985 1990 1995 2000 2005 2010 2015
Normalized Residential PPI
Normalized GDP position
0%
1%
2%
3%
4%
5%
6%
7%
8%
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1987 1992 1997 2002 2007 2012
Normalized Residential PPI
SIBOR 12-month
Low interest rate environment helps to stimulate housing activities
The role of government policies on housing markets
Recent government policies to cool the housing markets:
Measure Date Description of cooling measure
1 Sep-09Removal of interest absorption scheme and interest only housing loans
2 Feb-10Introduction of seller’s stamp duty (SSD) and lowering of LTV ratio from 90% to 80%
3 Aug-10Increasing the holding period of SSD and lowering LTV for multiple house loans
4 Jan-11Increasing the holding period of SSD and lowering LTV for multiple house loans
5 Dec-11Introduction of Additional Buyer’s Stamp Duty (ABSD) for foreign entities and multiple home owners.
6 Oct-12Maximum tenure of all new residential property loans will be capped at 35 years
7 Jan-13Reduction in the Debt Servicing Ratio to 30%; ABSD rate raised.
8 Jun-13New debt servicing framework introduced by MAS.
9 Feb-18Top marginal buyer's stamp duty go up from 3% to 4% for residential properties worth over $1 million.
Hypothesis test estimating the effectiveness of the cooling measures:
Decomposing housing prices: Economic variables
How effective were the cooling measures?
Cooling Measures 6 & 7:• Significantly reduced transaction
amounts and volumes• Restriction of loan tenures for
residential properties• Increase in ABSD, tightening of LTV
etcCooling Measure 3:• Increasing SSD holding period from
1 to 3 years• For buyers with one or more
outstanding housing loans, (i) an increase in minimum cash payment and (ii) decrease in LTV ratio
Decomposing housing prices: Economic variables
Are Singapore housing prices sensitive to foreign demand forces?
Top 10 nationalities of purchasers:Hypothesis test confirming the role of foreign demand and foreign exchange rate strengthening on transacted volume:
Decomposing housing prices: Speculation variables
Where to from here? Forecasting Singapore’s future housing markets
Projecting the direction of Singapore’s housing market
Long term expectations:
Outlook Optimistic Best Estimate Pessimistic
On demographic factors ++ +/- --
On economic factors ++ +/- -
On government policies +/- +/- +/-
Where to from here? Forecasting Singapore’s future housing markets
Best estimates, optimistic and pessimistic outlooks for Singapore
Outlook Optimistic Best Estimate Pessimistic
On demographic factors
Fertility rate improves through a variety of successful promotionalcampaigns. Population growth is also driven by the close-to-double digit growth of non-residents.
Fertility rate improves gradually, with population growth only driven by a positive net immigration. Real income per household to improve with improved education.
Fertility rate remains low at current rate of 1.2, leading to ageing population.
On economic factors
More jobs and output are created as part of the nation’s push towards embracing new technologies and becoming a global hub for IP.
Interest rates continues to track Fed rates, but remains under 200 bps.
Rising interest rates to track Fed rates. Coupled with trade instability in the region, which Singapore is adversely affected.
On government policies
Continue to enforce countercyclical measures, to maintain housing prices close to their fundamental values.
Continue to enforce countercyclical measures, to maintain housing prices close to their fundamental values
Continue to enforce countercyclical measures, to maintain housing prices close to their fundamental values
Singapore Actuarial Society Health and Retirement Conference 2018
Section 2: Modelling house prices – A relative value perspective
Aggregating the information out there…
Main Data: URA
Transaction data:• 280k records• 10yrs from 2008• Information: Price, size,
tenure, completion date, address, dwelling type
Economic Data: Varied Sources
Macroeconomic data:• To control for differences
in economic environment• Information: GDP, SIBOR,
Vacancy Rates, Inflation
Demographics Data: Census
Census data:• To control for differences
in demographics• Information: Population
density, Household Incomes
Geospatial Data: Googlemaps
Geospatial data:• Engineer additional
features that describe geospatial differences
• Information: Distances to key landmarks
Mer
gin
g &
Tra
nsf
orm
atio
ns
Cleaned data:
Insights from data: Part 1 – Geographical variables
Location, location, location – What is in a location?
Insights from data: Part 1 – Geographical variables
Distance to top school – A proxy for kiasuism
Source: www.kiasuparents.com
Encoding the ‘top school’ variable
Source: MOE
Increasing your baby’s chance of a better future
Insights from data: Part 1 – Geographical variables
Distance to train stations – How do we get to work/school?
Insights from data: Part 2 – Property characteristic variables
High rise vs low rise – No surprises hereSingapore River, View from the top Singapore River, View from the ground
Insights from data: Part 3 – Age and tenure variables
Property Type – Leasehold vs Freehold
Unveiling our toolbox – Learning Algorithms
Item How does it work? Main Tuning Parameters
Linear Approaches
OLS Linear RegressionEstimate a linear relationship between response and explanatory variables
None
LASSOExtends OLS linear regression with a penalty to complexity of the model
Regularisation to force non-essential coefficients to zero
Tree Based Learning Approaches
Decision TreeNon-parametric and non-linear approach partitioning to explain the response variables
Depth, Minimum partition size, Number of features
Bootstrap Aggregation (Bagging Trees)
Ensemble approach that creates multiple trees on different samples and average the means across each sample
Number of trees, tree depth
Random ForestEnsemble approach similar to bagging, but only a subset of features are considered for node splitting
Number of trees, tree depth, number of features considered for splitting
Gradient Boosted Machine
Ensemble approach that fits trees and learns from the areas that the model is fitting poorly
Number of trees, tree depth, learning rate
Additional Non-linear Learning Approaches
Support Vector Regression
Non-parametric and non-linear approach by fitting a hyperplane to separate the continuous variables
Choice of kernel, penalty factor
K-Nearest-NeighbourNon-parametric and non-linear approach by taking the average values of the k nearest neighbours
k
Considered statistical learning approaches:
Developing our statistical model
Visualising model accuracy (Out of sample):
Developing our statistical model
Ranking our key variables
Bring both models together
How can we price a house transaction in 2018Q4?
Step 1: Obtain transactional property characteristics of the key house (ie address/floor)
Step 2: Estimate economic landscape of 2018Q4 using the macroeconomic model
Step 3: Apply both transactional variables and macroeconomic forecasts to transactional model
Final valuation that uses (1) a statistical learning model and (2) a traditional approach to forecasting macroeconomic conditions
Singapore Actuarial Society Health and Retirement Conference 2018
Questions and Answers